International Journal of Pure and Applied Mathematics

Volume 85 No. 3 2013, 583-592

ISSN: 1311-8080 (printed version); ISSN: 1314-3395 (on-line version)
url: [Link]
doi: [Link]
AP
[Link]

A COMPARISON OF PERFORMANCE OF RESIDUAL

CONTROL CHARTS FOR TREND STATIONARY
AR(p) PROCESSES

Yupaporn Areepong
Department of Applied Statistics
Faculty of Applied Science
King Mongkut’s University of Technology North Bangkok
Bangkok, 10800, THAILAND

Abstract: The main goal of this paper is to implement Statistical Process

Control (SPC) with Residual Control Charts. In real life applications, such as
in industries or manufacturing, many processes are described by autocorrelated
observations. In addition to autocorrelation, some types of industrial processes
also exhibit a particular kind of trend behavior. A typical example of process
with autocorrelations arises in chemical processes. One of the main character-
istics of a control chart are the Average Run Length (ARL0 ) (mean of false
alarm times) and the Average Delay time (ARL1 ) (mean delay of true alarm
times). The main parameter observed for a given process is the mean shift. If a
process has no mean shift, then the ARL0 should be sufficiently large. On the
opposite side, when a mean shift occurs the ARL1 should be small in order to
indicate the mean shift quickly. We compared the performance of the Shewhart
- x̄, EWMA, and GMA residual control charts for autocorrelation observations
with upward/downward linear trend. These comparisons are made for different
magnitudes of the mean shift and various levels of autocorrelation.

AMS Subject Classification: 60A05

Key Words: Shewhart - x̄ chart, exponentially weighted moving average con-
c 2013 Academic Publications, Ltd.

Received: February 27, 2013 url: [Link]
584 Y. Areepong

trol chart (EWMA), geometric moving average control chart (GMA), average
run length (ARL)

1. Introduction

Statistical Process Control (SPC) techniques are widely used for monitoring
and improving quality of production in many areas of practical applications
application, for example, industry and manufacturing, finance and economics,
epidemiology and health care, environmental sciences and for many other fields
see, Borror et al. [3]; Stoumbos and Reynolds [7]. Traditional control charts are
evaluated under the assumption that observations are independent and identi-
cally distributed (i.i.d.) and in most cases normally distribution. However, in
real applications, many charts have autocorrelated observations. As an exam-
ple, the processes implemented in a control chart to detect computer intrusions
in network systems, or the manufacture of food, chemicals, paper and other
wood products. As a results, recently, has been an increasing research interest
to study the effect of autocorrelation on the performance of SPC charts.
It has been observed that, the main effect of autocorrelation in the process
data within a SPC schemes for a traditional chart the Average Run Length
of in-control process may be more shorter than the intended one (see Johnson
and Bagshaw [5]). Processes with serial correlation in data also need to be
monitored by appropriate control charts.
One of the most widely approach is to fit the autocorrelated data of the
observed process into a time series model, such that forecasts of each observation
can be made using the previous observations, and apply the fitted data to the
residuals of traditional control charts (see e.g. Harris and Ross [6], Lu and
Reynolds [4]).
Johnson and Bagshaw [5] derived approximate run length distribution for
the Cumulative Sum Control Chart (CUSUM) chart when the process follows
a Moving Average model of order one, MA(1), or an Autoregressive Process of
order one, AR(1). Harris and Ross [6] discussed the effect of autocorrelation on
the performance of Exponentially Weighted Moving Average (EWMA) charts
and CUSUM, and found that the Average Run Lengths of these charts were
sensitive to the presence of autocorrelation.
Recently, Karaoglan and Bayhan [1] compared performance of Shewhart,
CUSUM EWMA and GMA residual control charts for increasing trend of an
AR(1) process. In this paper, we study the residual control charts for AR(p)
process which includes upwards and downward linear trend. We also compare
A COMPARISON OF PERFORMANCE OF RESIDUAL CONTROL... 585

the performance of the Shewhart - x̄, EWMA, and GMA residual control charts
for autocorrelation observations at different magnitudes of the mean shift and
various levels of autocorrelation.

2. Control Charts and Their Properties

The Average Run Length is an important characteristic for any SPC chart.
It is defined as the expectation of the time before the control chart gives a
false alarm that an in-control process has run out-of-control (ARL0 ). A second
important characteristic for a SPC chart is the Average Delay time (ARL1 ).
To design and efficient chat the ARL0 should be large and the ARL1 should
be small.
Let Eθ (.) denote the expectation that the change-point occurs in a random
observations generated from a distribution function F (x, µ, σ 2 ). At one point
in time θ the distribution mean is change from µ = µ0 to µ = µ1 , where θ ≤ ∞.
A typical condition imposed on an ARL0 is

Eθ (τ ) = T, θ = ∞ (in-control state), (1)

where T is given (usually large). For a given distribution function and chart,
this condition determine the choices of UCL and LCL and the typical practical
constraint is ARL0 = T .
A typical definition of the ARL1 is

ARL1 = E1 (τ |τ ≥ 1), (2)

means that the change point occurs at θ = 1. One could expect that a sequential
control chart has a near optimal performance if its ARL1 is close to a minimal
value. There are many other criteria that could be used for designing optimal
SPC charts.
Let xt be autocorrelated observations, then the residual forms a time series
model of xt defined as
et = xt − x̂t , (3)
where x̂t is the prediction of xt at time t. Residual control charts are construct
based on et depending on the traditional charts used. For a Shewhart residual
chart,the 3σ upper and lower control limits are defined by

U CL = ē + 3σe , LCL = ē − 3σe ,

where ē is the center line of the chart with standard deviation σe .

586 Y. Areepong

The alarm time for the Shewhart residual chart is given by:

τ = inf{t > 0 : et > U CL or et < LCL}.

An EWMA chart is defined by a recursive formula

Yt = (1 − λ)Yt−1 + λet , t = 1, 2, ...

where λ ∈ (0, 1) is a smoothing parameter and Yt is the weighted average

between current and previous observations at t. The target mean is supposed
to be steady and the initial value et is usually chosen to be the process mean
µ0 . The upper control limit of the EWMA chart is the following:
r
λ
U CL = ē + Lσe ,
2−λ
and the lower control limit is:
r
λ
LCL = ē − Lσe ,
2−λ
where L is a constant to be chosen. The process will be declared to be in an
out-of-control state when Yt > U CL or Yt < LCL . The alarm time for the
EWMA is then given by

τ = inf{t > 0 : Yt > U CL or Yt < LCL}.

A Geometric Moving Average(GMA) control chart is defined by the following

statistics (Karaoglan and Bayhan [2]):

Yt = (1 − λ)Yt−1 + λet , t = 1, 2, ...

The upper control limit of the GMA chart is the following:

U CL = ē + L(R̄/d2 )[λ/(2 − λ)]1/2 ,

and the lower control limit is:

LCL = ē − L(R̄/d2 )[λ/(2 − λ)]1/2 ,

where L is a constant to be chosen, R̄ is an average of range and d2 is the mean

of the distribution of the relative range which can be referred from statistical
A COMPARISON OF PERFORMANCE OF RESIDUAL CONTROL... 587

tables. The process will be declared to be in an out-of-control state when

Yt > U CL or Yt < LCL . The alarm time for the GMA is then given by

τ = inf{t > 0 : Yt > U CL or Yt < LCL}.

In general, a p th - order AR model is defined as follow

zt = c + φ1 zt−1 + φ2 zt−2 + ... + φp zt−p + et

where c is a constant, φj is j th autoregressive parameters and et is the error

term of time t. There are specific restriction for different cases:

when p = 1, −1 < φ1 < 1 and

when p = 2, −1 < φ2 < 1, φ1 + φ2 < 1, φ2 − φ1 < 1.

Let Zt be a linear trend represented by:

Zt = zt + dt

where d is the trend slope in term of t.

3. Comparison of Performance of Control Charts

In this section, the numerical results for ARL0 and ARL1 for Shewhart, EWMA
and GMA residual charts are presented. Table 1 shows a comparison of ARL
for AR(1) process with values obtained via simulation.

The parameter values for Shewhart, EWMA and GMA residual charts were
chosen by setting the desired ARL0 = 370, φ1 = 0.75, the value of the in-control
parameter µ = 0 and the shift parameter δ ⊆ [0.5; 4].
For the EWMA and GMA procedures, the parameter values λ = 0.05 and
U CL = 2.504 and 2.613 respectively. The results shown that for d = 0.5, the
EWMA residual chart have a better performance than the other charts when
the value of the parameter δ had a small and moderate shifts.
For example, when δ = 1 the EWMA chart shows the best performance
because it gives the minimum AD value (17.736). However, the performance of
the Shewhart residual chart is superior to EWMA and GMA charts when δ > 2,
i.e., when the shift is large. The results for d = −0.5 are in good agreement
with the results for the results d = 0.5.
588 Y. Areepong

Table 1: Comparison of ARL for AR(1) on Shewhart EWMA and GMA

residual charts when given ARL0 =370 and φ1 = 0.75.

ARL
d δ Shewhart EWMA GMA
0.5 0 369.459 371.336 371.222
0.5 177.452 35.568∗ 35.649
1 58.875 17.736∗ 17.740
1.5 25.941 13.583∗ 13.589
2 15.553 11.961∗ 11.966
2.5 11.034∗ 11.324 11.323
3 8.262∗ 11.155 11.159
3.5 5.416∗ 9.468 9.582
4 3.128∗ 6.412 6.658
-0.5 0 369.899 370.650 371.398
0.5 166.136 27.438∗ 27.460
1 47.349 10.259∗ 10.272
1.5 14.205 5.951∗ 5.954
2 4.719 4.192∗ 4.194
2.5 2.089∗ 3.278 3.279
3 1.279∗ 2.716 2.719
3.5 1.254 ∗ 2.584 2.596
4 1.106∗ 2.109 2.201

Table 2 gives a comparison of ARL for AR(1) process. The parameter

values for Shewhart, EWMA and GMA residual charts were chosen by setting
the desired ARL0 = 370, φ1 = −0.5, results for ARL0 and ARL1 for Shewhart,
EWMA and GMA residual charts presented in a similar manner to Table 1.
Table 3 shows a comparison of ARL for AR(2) process. The parameter
values for Shewhart, EWMA and GMA residual charts were chosen by setting
the desired ARL0 = 370, φ1 = 0.75, φ2 = −0.5, the value of the in-control
parameter µ = 0 and the shift parameter δ ⊆ [0.5; 4].
For the EWMA and GMA procedures, the parameter values λ = 0.05 and
U CL = 2.204 and 2.137 respectively. The results shown that for d = 0.5, the
EWMA residual chart performs better than the other charts when the value of
the parameter δ had a small and moderate shifts.
For example, when δ = 0.5 the EWMA chart shows the best performance
A COMPARISON OF PERFORMANCE OF RESIDUAL CONTROL... 589

Table 2: Comparison of ARL for AR(1) on Shewhart EWMA and GMA

residual charts when given ARL0 =370 and φ1 = −0.5.

ARL
d δ Shewhart EWMA GMA
0.5 0 370.002 369.073 369.532
0.5 83.315 15.467∗ 15.469
1 16.349 7.248∗ 7.249
1.5 5.312 5.181∗ 5.183
2 2.742 4.258∗ 7.259
2.5 2.137∗ 3.755 3.757
3 1.847∗ 2.747 2.749
3.5 1.467 ∗ 2.516 2.524
4 1.108 ∗ 2.018 2.124
-0.5 0 371.817 370.777 370.221
0.5 85.748 16.132∗ 16.151
1 18.016 7.709∗ 7.713
1.5 5.613 5.493∗ 5.496
2 2.794∗ 4.471 4.472
2.5 2.166∗ 3.912 3.916
3 1.736∗ 2.487 2.487
3.5 1.358 ∗ 2.195 2.198
4 1.193 ∗ 1.872 1.948

because the AD have a minimal value (39.773). However, the performance of

the Shewhart residual chart is superior to EWMA and GMA charts when δ > 2,
i.e., when the shift is large. The results for d = −0.5 are in good agreement
with the results for the results d = 0.5.
In Table 4, the comparison of ARL for an AR(2) process are presented.
The parameter values for Shewhart, EWMA and GMA residual charts were
chosen by setting the desired ARL0 = 370, φ1 = 0.5, φ2 = 0.25, the results for
d = 0.5 shown that the EWMA residual chart have a best performance for small
and moderate shifts but Shewhart residual chart perform better than EWMA
and GMA charts when δ > 3. For d = −0.5, the performance of the EWMA
residual chart is superior to Shewhart and GMA charts when 0.5 ≤ δ ≤ 2 but
the performance of the Shewhart residual chart is superior to EWMA and GMA
charts when δ > 2.
590 Y. Areepong

Table 3: Comparison of ARL for AR(2) on Shewhart EWMA and GMA

residual charts when given ARL0 =370, φ1 = 0.75 and φ2 = −0.5.

ARL
d δ Shewhart EWMA GMA
0.5 0 370.609 369.508 370.983
0.5 207.821 39.773∗ 39.874
1 77.479 15.934∗ 15.942
1.5 29.164 9.756∗ 9.760
2 11.066 6.941∗ 6.945
2.5 4.370∗ 5.270 5.272
3 2.487∗ 4.171 4.171
3.5 2.295 ∗ 3.465 3.601
4 1.954∗ 2.841 2.934
-0.5 0 370.757 370.544 370.291
0.5 188.674 34.929∗ 34.902
1 58.855 13.148∗ 13.188
1.5 18.051 7.902∗ 7.903
2 5.613 5.535∗ 5.536
2.5 2.711∗ 4.233 4.253
3 2.139∗ 3.523 3.521
3.5 1.782 ∗ 3.185 3.142
4 1.428∗ 2.683 2.752

4. Conclusion

We compared the effectiveness of the Shewhart - x̄, EWMA, and GMA residual
control charts for autocorrelation observations with upward and downward lin-
ear trend. The comparison of the control charts was based on the Average Run
Length (ARL0 )and the Average Delay time (ARL1 ) criteria. For all orders of
an AR(p) process analyzed in this research, we shown that the performance of
EWMA charts is superior to the Shewhart - x̄ and GMA residual charts for
small shifts, however the performance of Shewhart - x̄ residual chart is superior
to EWMA and GMA residual charts for large shifts.
A COMPARISON OF PERFORMANCE OF RESIDUAL CONTROL... 591

Table 4: Comparison of ARL for AR(2) on Shewhart EWMA and GMA

residual charts when given ARL0 =370, φ1 = 0.5 and φ2 = 0.25.

ARL
d δ Shewhart EWMA GMA
0.5 0 370.499 370.068 370.686
0.5 166.291 33.590∗ 33.784
1 54.400 17.441∗ 17.831
1.5 25.126 13.676∗ 13.774
2 15.949 12.012∗ 12.014
2.5 12.473 11.117∗ 11.118
3 10.717 10.701∗ 10.703
3.5 8.542∗ 9.521 9.698
4 6.294 ∗ 8.293 8.649
-0.5 0 371.929 370.795 370.006
0.5 160.710 26.162∗ 26.347
1 43.910 10.655∗ 10.662
1.5 13.773 6.780∗ 6.782
2 5.401 5.170∗ 5.184
2.5 3.099∗ 4.321 4.322
3 2.384∗ 3.796 3.798
3.5 2.190∗ 2.691 2.721
4 1.755 ∗ 2.019 2.251

5. Acknowledgments

The author would like to give special thanks to [Link] Mititelu for careful
proof-reading and suggestions. The author would also like to express her grat-
itude to Faculty of Science, King Mongkut’s University of Technology, North
Bangkok, Thailand for a supporting research grant of No: 5645105.

References

[1] A.D. Karaoglan and G.M. Bayhan, Performance Comparison of Residual

Control Charts for Trend Stationary First Order Autoregressive Processes,
Gazi University Journal of Science, 24 (2011), 329-339.
592 Y. Areepong

[2] A.D. Karaoglan and G.M. Bayhan, ARL performance of residual control
charts for trend AR(1) process: A case study on peroxide values of stored
vegetable oil, Scientific Research and Essays, 7 (2012), 1405-1414.

[3] C.M. Borror, D.C. Montgomery and G.C. Runger, Robustness of the
EWMA Control Chart to Non-normality, Journal of Quality Technology,
31 (1999), 309-316.

[4] C.W. Lu and M.R. Reynolds, EWMA control charts for monitoring the
mean of autocorrelated processes, J Qual Technol, 31 (1999), 166-188.

[5] R. A. Johnson and M. Bagshaw, The effect of serial correlation on the

performance of CUSUM tests, Technometrics, 16 (1974), 103-112.

[6] T. J. Harris and W. H. Ross, Statistical process control procedures for

correlated observations, Canadian Journal of Chemical Engineering, 69,
(1991), 48-57.

[7] Z.G. Stoumbos and M.R. Reysnolds, Robustness to non-normality and

autocorrelation of individuals control charts for monitoring the process
mean and variance, Journal of Statistical Computation and Simulation,66
(2000), 145-187.